Effective WBC Segmentation Using Hybrid Loss

The performance of computer-aided blood related disease detection depends on accuracy of white blood cell segmentation. The segmentation task is very challenging mainly due to data imbalance problem. To address the problem of data imbalance a hybrid loss is proposed which focus more on pixels which are mispredicted and less in number. The hybrid loss achieves a good trade off between recall and precision by reducing the counts of false positive and false negative instances. The hybrid loss helped in achieving F1-score of 0.936, 0.868, and 0.943 using UNet, attention based UNet, and attention based UNet having the residual block, respectively.